WangchanLION-v3

WangchanLION-v3 is an 8-billion-parameter Thai language model released around November 2025 as a collaboration between AI Singapore, VISTEC, and SCB10X. It was pretrained on 47 billion high-quality Thai tokens, filtered down from 100 billion, with the dataset released on AI Singapore's Hugging Face alongside the weights.

Unlike other Thai pretraining efforts that focus on releasing models, the project openly shares the data cleaning pipeline experiments and SFT results, aiming to show what it takes to build a high-quality Thai pretraining corpus with transparent, reproducible methods. It continues the WangchanLion line of SEA-LION-based Thai instruction models. Check the exact license terms on the model card before commercial use, as the SEA-LION family carries research-oriented licensing.